Dual-Stream Multimodal Learning: Synergizing 3D EfficientNet and XGBoost with Interpretability for Early Alzheimers Detection

The increasing worldwide prevalence of Alzheimer's disease (AD) prompts a pressing need for new diagnostic systems that address the limitations of traditional (clinical) methods, that are frequently subjective or cost prohibitive, while only being useful in the most advanced stages of AD. In this paper, we present an innovative multimodal deep learning system designed to improve the early and accurate prediction of AD by combining structural neuroimaging data (T1 weighted MRI scans) with structured clinical data (cognition scores and demographics). Our model functions using a two-pronged approach: one pathway utilizes a 3D EfficientNet to examine volumetric MRI data and identify cerebral atrophy; while the second pathway utilizes an XGBoost Classifier (Gradient Boosted Trees) to collect predictive patterns from tabular clinical data. This dual modality provides a "double verification" approach to diagnostic confidence along with greater resilience to changes in clinical features. A key contribution of our work is the dual-modality interpretability engine which utilizes 3D Grad-CAM visualizations and SHapley Additive exPlanations (SHAP) force plots to explain which clinical features has the greatest influence on the model. We anticipate that the resulting framework will outperform singlemodality approaches in distinguishing between normal cognition (NC), mild cognitive impairment (MCI), and AD cases.